My current research focusses on methods for A/B test context validation in dynamic environments for web applications.

Data-driven and continuous development and deployment of modern web applications depend critically on registering changes as fast as possible, paving the way for short innovation cycles. A/B testing is a good candidate for comparing the performance of different versions. Despite the widespread use of A/B tests, there is little research on how to assert the validity of such tests. Even small changes in the application’s user base, hard- or software stack not related to the variants under test can transform on possibly hidden paths into significant disturbances of the overall performance of an A/B test and, hence, invalidate such a test. Therefore, the highly dynamic server and client run-time environments of modern web applications make it difficult to correctly assert the validity of an A/B test.

I will list accepted papers here.

During my time at Volkswagen, I filed the following patent: Vehicle Internet API
In a sentence: An API for creating car-centric Apps to build an AppStore for car manufacturers.


I did my Ph.D. at the University of Hanover under the supervision of Prof. Jörg Hähner and Prof. Monika Sester. I started in August 2007 and submitted my thesis in July 2010. It took about three years of work to complete the Ph.D., which is not too bad for a 100% research program (only a minimum of lecturing was required).

People often ask me how to become a data scientist. The Ph.D. was my way of doing that. By starting a Ph.D. in systems engineering, I knew that I would have the opportunity to work on interesting problems. The Ph.D. opened me a door to a fascinating job market due to the boom in engineering complex data-driven systems, which wasn’t something I expected. All I was hoping to achieve was being qualified to work on more interesting stuff than plain software engineers.

Broadly speaking, the topics of the Ph.D. were in the areas of real-time configuration of active cameras. Although seeming to have not much in common with big data analytics, active cameras implement the map/reduce paradigm by pre-processing vast amounts of image data locally to send only abstract information. Furthermore, they have to analyze jointly (a distributed reduce step) which targets to observe and predict their optimal positions, which is very similar to methods used for predictive analytics for real-time processing. Additionally, computer vision is full of statistics. I’m planning to eventually document my research journey and point out further similarities between both topics through a series of posts.

The title of my thesis is "Dynamic reconfiguration algorithms for Active Camera Networks". The short, human-friendly abstract is:

Active Camera Networks consist of autonomous vehicles - each one equipped with a visual sensor - communicating wirelessly with each other to perform surveillance tasks in a collaborative way. This thesis is devoted to the problem of wide-area target acquisition of moving targets in a surveillance area. It addresses application scenarios where events unfold over a large geographic area, and close-up views have to be acquired for biometric tasks such as face detection. The main problem is to coordinate numerous cameras to reach a system behavior that only one capture of each target is acquired.

The main publications that resulted from my Ph.D. work are as follows (before I married my last name was Wittke):